When people talk about non-determinism of floating point, what they usually mean is non-associativity, that is (x+y)+z may not be exactly equal to x+(y+z).
Good example of this, in Python 3:
>>> (0.1 + 0.2) + 0.3
>>> 0.1 + (0.2 + 0.3)
It's not an optimisation if it changes the result! And if you use non-standard flags that's your problem.
MP3 is an optimization of WAV, yet it changes the result.
Some applications are ok with reducing precision of calculations because they are not sensitive enough to small inaccuracies or they take effort to control inaccuarcies.
For example, graphics applications are typically heavy in FP calculations and yet they tend to not care much about precision and much more about performance. For those applications reducing accuracy for slight performance increase is likely win.